6961690: load oops from constant table on SPARC

Summary: oops should be loaded from the constant table of an nmethod instead of materializing them with a long code sequence.
Reviewed-by: never, kvn

10 files changed, 501 insertions, 45 deletions

diff --git a/src/share/vm/opto/c2_globals.hpp b/src/share/vm/opto/c2_globals.hpp
index 2bdc1a04a..fe516dff5 100644
--- a/src/share/vm/opto/c2_globals.hpp
+++ b/src/share/vm/opto/c2_globals.hpp
@@ -284,6 +284,9 @@
   develop(bool, SparcV9RegsHiBitsZero, true,                                \
           "Assume Sparc V9 I&L registers on V8+ systems are zero-extended") \
                                                                             \
+  product(bool, UseRDPCForConstantTableBase, false,                         \
+          "Use Sparc RDPC instruction for the constant table base.")        \
+                                                                            \
   develop(intx, PrintIdealGraphLevel, 0,                                    \
           "Print ideal graph to XML file / network interface. "             \
           "By default attempts to connect to the visualizer on a socket.")  \
diff --git a/src/share/vm/opto/compile.cpp b/src/share/vm/opto/compile.cpp
index 972cc34ae..0ca15e168 100644
--- a/src/share/vm/opto/compile.cpp
+++ b/src/share/vm/opto/compile.cpp
@@ -75,6 +75,18 @@
 # include "adfiles/ad_zero.hpp"
 #endif
 
+
+// -------------------- Compile::mach_constant_base_node -----------------------
+// Constant table base node singleton.
+MachConstantBaseNode* Compile::mach_constant_base_node() {
+  if (_mach_constant_base_node == NULL) {
+    _mach_constant_base_node = new (C) MachConstantBaseNode();
+    _mach_constant_base_node->add_req(C->root());
+  }
+  return _mach_constant_base_node;
+}
+
+
 /// Support for intrinsics.
 
 // Return the index at which m must be inserted (or already exists).
@@ -432,13 +444,14 @@ void Compile::print_compile_messages() {
 }
 
 
-void Compile::init_scratch_buffer_blob() {
-  if( scratch_buffer_blob() != NULL )  return;
+void Compile::init_scratch_buffer_blob(int const_size) {
+  if (scratch_buffer_blob() != NULL)  return;
 
   // Construct a temporary CodeBuffer to have it construct a BufferBlob
   // Cache this BufferBlob for this compile.
   ResourceMark rm;
-  int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size);
+  _scratch_const_size = const_size;
+  int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
   BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size);
   // Record the buffer blob for next time.
   set_scratch_buffer_blob(blob);
@@ -455,9 +468,19 @@ void Compile::init_scratch_buffer_blob() {
 }
 
 
+void Compile::clear_scratch_buffer_blob() {
+  assert(scratch_buffer_blob(), "no BufferBlob set");
+  set_scratch_buffer_blob(NULL);
+  set_scratch_locs_memory(NULL);
+}
+
+
 //-----------------------scratch_emit_size-------------------------------------
 // Helper function that computes size by emitting code
 uint Compile::scratch_emit_size(const Node* n) {
+  // Start scratch_emit_size section.
+  set_in_scratch_emit_size(true);
+
   // Emit into a trash buffer and count bytes emitted.
   // This is a pretty expensive way to compute a size,
   // but it works well enough if seldom used.
@@ -476,13 +499,20 @@ uint Compile::scratch_emit_size(const Node* n) {
   address blob_end   = (address)locs_buf;
   assert(blob->content_contains(blob_end), "sanity");
   CodeBuffer buf(blob_begin, blob_end - blob_begin);
-  buf.initialize_consts_size(MAX_const_size);
+  buf.initialize_consts_size(_scratch_const_size);
   buf.initialize_stubs_size(MAX_stubs_size);
   assert(locs_buf != NULL, "sanity");
-  int lsize = MAX_locs_size / 2;
-  buf.insts()->initialize_shared_locs(&locs_buf[0],     lsize);
-  buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize);
+  int lsize = MAX_locs_size / 3;
+  buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
+  buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
+  buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
+
+  // Do the emission.
   n->emit(buf, this->regalloc());
+
+  // End scratch_emit_size section.
+  set_in_scratch_emit_size(false);
+
   return buf.insts_size();
 }
 
@@ -516,10 +546,13 @@ Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr
                   _orig_pc_slot(0),
                   _orig_pc_slot_offset_in_bytes(0),
                   _has_method_handle_invokes(false),
+                  _mach_constant_base_node(NULL),
                   _node_bundling_limit(0),
                   _node_bundling_base(NULL),
                   _java_calls(0),
                   _inner_loops(0),
+                  _scratch_const_size(-1),
+                  _in_scratch_emit_size(false),
 #ifndef PRODUCT
                   _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
                   _printer(IdealGraphPrinter::printer()),
@@ -783,6 +816,7 @@ Compile::Compile( ciEnv* ci_env,
     _failure_reason(NULL),
     _code_buffer("Compile::Fill_buffer"),
     _has_method_handle_invokes(false),
+    _mach_constant_base_node(NULL),
     _node_bundling_limit(0),
     _node_bundling_base(NULL),
     _java_calls(0),
@@ -2862,3 +2896,207 @@ Compile::TracePhase::~TracePhase() {
     _log->done("phase nodes='%d'", C->unique());
   }
 }
+
+//=============================================================================
+// Two Constant's are equal when the type and the value are equal.
+bool Compile::Constant::operator==(const Constant& other) {
+  if (type()          != other.type()         )  return false;
+  if (can_be_reused() != other.can_be_reused())  return false;
+  // For floating point values we compare the bit pattern.
+  switch (type()) {
+  case T_FLOAT:   return (_value.i == other._value.i);
+  case T_LONG:
+  case T_DOUBLE:  return (_value.j == other._value.j);
+  case T_OBJECT:
+  case T_ADDRESS: return (_value.l == other._value.l);
+  case T_VOID:    return (_value.l == other._value.l);  // jump-table entries
+  default: ShouldNotReachHere();
+  }
+  return false;
+}
+
+// Emit constants grouped in the following order:
+static BasicType type_order[] = {
+  T_FLOAT,    // 32-bit
+  T_OBJECT,   // 32 or 64-bit
+  T_ADDRESS,  // 32 or 64-bit
+  T_DOUBLE,   // 64-bit
+  T_LONG,     // 64-bit
+  T_VOID,     // 32 or 64-bit (jump-tables are at the end of the constant table for code emission reasons)
+  T_ILLEGAL
+};
+
+static int type_to_size_in_bytes(BasicType t) {
+  switch (t) {
+  case T_LONG:    return sizeof(jlong  );
+  case T_FLOAT:   return sizeof(jfloat );
+  case T_DOUBLE:  return sizeof(jdouble);
+    // We use T_VOID as marker for jump-table entries (labels) which
+    // need an interal word relocation.
+  case T_VOID:
+  case T_ADDRESS:
+  case T_OBJECT:  return sizeof(jobject);
+  }
+
+  ShouldNotReachHere();
+  return -1;
+}
+
+void Compile::ConstantTable::calculate_offsets_and_size() {
+  int size = 0;
+  for (int t = 0; type_order[t] != T_ILLEGAL; t++) {
+    BasicType type = type_order[t];
+
+    for (int i = 0; i < _constants.length(); i++) {
+      Constant con = _constants.at(i);
+      if (con.type() != type)  continue;  // Skip other types.
+
+      // Align size for type.
+      int typesize = type_to_size_in_bytes(con.type());
+      size = align_size_up(size, typesize);
+
+      // Set offset.
+      con.set_offset(size);
+      _constants.at_put(i, con);
+
+      // Add type size.
+      size = size + typesize;
+    }
+  }
+
+  // Align size up to the next section start (which is insts; see
+  // CodeBuffer::align_at_start).
+  assert(_size == -1, "already set?");
+  _size = align_size_up(size, CodeEntryAlignment);
+
+  if (Matcher::constant_table_absolute_addressing) {
+    set_table_base_offset(0);  // No table base offset required
+  } else {
+    if (UseRDPCForConstantTableBase) {
+      // table base offset is set in MachConstantBaseNode::emit
+    } else {
+      // When RDPC is not used, the table base is set into the middle of
+      // the constant table.
+      int half_size = _size / 2;
+      assert(half_size * 2 == _size, "sanity");
+      set_table_base_offset(-half_size);
+    }
+  }
+}
+
+void Compile::ConstantTable::emit(CodeBuffer& cb) {
+  MacroAssembler _masm(&cb);
+  for (int t = 0; type_order[t] != T_ILLEGAL; t++) {
+    BasicType type = type_order[t];
+
+    for (int i = 0; i < _constants.length(); i++) {
+      Constant con = _constants.at(i);
+      if (con.type() != type)  continue;  // Skip other types.
+
+      address constant_addr;
+      switch (con.type()) {
+      case T_LONG:   constant_addr = _masm.long_constant(  con.get_jlong()  ); break;
+      case T_FLOAT:  constant_addr = _masm.float_constant( con.get_jfloat() ); break;
+      case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
+      case T_OBJECT: {
+        jobject obj = con.get_jobject();
+        int oop_index = _masm.oop_recorder()->find_index(obj);
+        constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
+        break;
+      }
+      case T_ADDRESS: {
+        address addr = (address) con.get_jobject();
+        constant_addr = _masm.address_constant(addr);
+        break;
+      }
+      // We use T_VOID as marker for jump-table entries (labels) which
+      // need an interal word relocation.
+      case T_VOID: {
+        // Write a dummy word.  The real value is filled in later
+        // in fill_jump_table_in_constant_table.
+        address addr = (address) con.get_jobject();
+        constant_addr = _masm.address_constant(addr);
+        break;
+      }
+      default: ShouldNotReachHere();
+      }
+      assert(constant_addr != NULL, "consts section too small");
+      assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
+    }
+  }
+}
+
+int Compile::ConstantTable::find_offset(Constant& con) const {
+  int idx = _constants.find(con);
+  assert(idx != -1, "constant must be in constant table");
+  int offset = _constants.at(idx).offset();
+  assert(offset != -1, "constant table not emitted yet?");
+  return offset;
+}
+
+void Compile::ConstantTable::add(Constant& con) {
+  if (con.can_be_reused()) {
+    int idx = _constants.find(con);
+    if (idx != -1 && _constants.at(idx).can_be_reused()) {
+      return;
+    }
+  }
+  (void) _constants.append(con);
+}
+
+Compile::Constant Compile::ConstantTable::add(BasicType type, jvalue value) {
+  Constant con(type, value);
+  add(con);
+  return con;
+}
+
+Compile::Constant Compile::ConstantTable::add(MachOper* oper) {
+  jvalue value;
+  BasicType type = oper->type()->basic_type();
+  switch (type) {
+  case T_LONG:    value.j = oper->constantL(); break;
+  case T_FLOAT:   value.f = oper->constantF(); break;
+  case T_DOUBLE:  value.d = oper->constantD(); break;
+  case T_OBJECT:
+  case T_ADDRESS: value.l = (jobject) oper->constant(); break;
+  default: ShouldNotReachHere();
+  }
+  return add(type, value);
+}
+
+Compile::Constant Compile::ConstantTable::allocate_jump_table(MachConstantNode* n) {
+  jvalue value;
+  // We can use the node pointer here to identify the right jump-table
+  // as this method is called from Compile::Fill_buffer right before
+  // the MachNodes are emitted and the jump-table is filled (means the
+  // MachNode pointers do not change anymore).
+  value.l = (jobject) n;
+  Constant con(T_VOID, value, false);  // Labels of a jump-table cannot be reused.
+  for (uint i = 0; i < n->outcnt(); i++) {
+    add(con);
+  }
+  return con;
+}
+
+void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
+  // If called from Compile::scratch_emit_size do nothing.
+  if (Compile::current()->in_scratch_emit_size())  return;
+
+  assert(labels.is_nonempty(), "must be");
+  assert((uint) labels.length() == n->outcnt(), err_msg("must be equal: %d == %d", labels.length(), n->outcnt()));
+
+  // Since MachConstantNode::constant_offset() also contains
+  // table_base_offset() we need to subtract the table_base_offset()
+  // to get the plain offset into the constant table.
+  int offset = n->constant_offset() - table_base_offset();
+
+  MacroAssembler _masm(&cb);
+  address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
+
+  for (int i = 0; i < labels.length(); i++) {
+    address* constant_addr = &jump_table_base[i];
+    assert(*constant_addr == (address) n, "all jump-table entries must contain node pointer");
+    *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
+    cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
+  }
+}
diff --git a/src/share/vm/opto/compile.hpp b/src/share/vm/opto/compile.hpp
index 9ec26ac70..36d09aca4 100644
--- a/src/share/vm/opto/compile.hpp
+++ b/src/share/vm/opto/compile.hpp
@@ -48,7 +48,10 @@ class ConnectionGraph;
 class InlineTree;
 class Int_Array;
 class Matcher;
+class MachConstantNode;
+class MachConstantBaseNode;
 class MachNode;
+class MachOper;
 class MachSafePointNode;
 class Node;
 class Node_Array;
@@ -139,6 +142,81 @@ class Compile : public Phase {
     trapHistLength = methodDataOopDesc::_trap_hist_limit
   };
 
+  // Constant entry of the constant table.
+  class Constant {
+  private:
+    BasicType _type;
+    jvalue    _value;
+    int       _offset;         // offset of this constant (in bytes) relative to the constant table base.
+    bool      _can_be_reused;  // true (default) if the value can be shared with other users.
+
+  public:
+    Constant() : _type(T_ILLEGAL), _offset(-1), _can_be_reused(true) { _value.l = 0; }
+    Constant(BasicType type, jvalue value, bool can_be_reused = true) :
+      _type(type),
+      _value(value),
+      _offset(-1),
+      _can_be_reused(can_be_reused)
+    {}
+
+    bool operator==(const Constant& other);
+
+    BasicType type()      const    { return _type; }
+
+    jlong   get_jlong()   const    { return _value.j; }
+    jfloat  get_jfloat()  const    { return _value.f; }
+    jdouble get_jdouble() const    { return _value.d; }
+    jobject get_jobject() const    { return _value.l; }
+
+    int         offset()  const    { return _offset; }
+    void    set_offset(int offset) {        _offset = offset; }
+
+    bool    can_be_reused() const  { return _can_be_reused; }
+  };
+
+  // Constant table.
+  class ConstantTable {
+  private:
+    GrowableArray<Constant> _constants;          // Constants of this table.
+    int                     _size;               // Size in bytes the emitted constant table takes (including padding).
+    int                     _table_base_offset;  // Offset of the table base that gets added to the constant offsets.
+
+  public:
+    ConstantTable() :
+      _size(-1),
+      _table_base_offset(-1)  // We can use -1 here since the constant table is always bigger than 2 bytes (-(size / 2), see MachConstantBaseNode::emit).
+    {}
+
+    int size() const { assert(_size != -1, "size not yet calculated"); return _size; }
+
+    void set_table_base_offset(int x)  { assert(_table_base_offset == -1, "set only once");                        _table_base_offset = x; }
+    int      table_base_offset() const { assert(_table_base_offset != -1, "table base offset not yet set"); return _table_base_offset; }
+
+    void emit(CodeBuffer& cb);
+
+    // Returns the offset of the last entry (the top) of the constant table.
+    int  top_offset() const { assert(_constants.top().offset() != -1, "constant not yet bound"); return _constants.top().offset(); }
+
+    void calculate_offsets_and_size();
+    int  find_offset(Constant& con) const;
+
+    void     add(Constant& con);
+    Constant add(BasicType type, jvalue value);
+    Constant add(MachOper* oper);
+    Constant add(jfloat f) {
+      jvalue value; value.f = f;
+      return add(T_FLOAT, value);
+    }
+    Constant add(jdouble d) {
+      jvalue value; value.d = d;
+      return add(T_DOUBLE, value);
+    }
+
+    // Jump table
+    Constant allocate_jump_table(MachConstantNode* n);
+    void         fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const;
+  };
+
  private:
   // Fixed parameters to this compilation.
   const int             _compile_id;
@@ -212,6 +290,11 @@ class Compile : public Phase {
   Node*                 _recent_alloc_obj;
   Node*                 _recent_alloc_ctl;
 
+  // Constant table
+  ConstantTable         _constant_table;        // The constant table for this compile.
+  MachConstantBaseNode* _mach_constant_base_node;  // Constant table base node singleton.
+
+
   // Blocked array of debugging and profiling information,
   // tracked per node.
   enum { _log2_node_notes_block_size = 8,
@@ -272,6 +355,8 @@ class Compile : public Phase {
   static int            _CompiledZap_count;     // counter compared against CompileZap[First/Last]
   BufferBlob*           _scratch_buffer_blob;   // For temporary code buffers.
   relocInfo*            _scratch_locs_memory;   // For temporary code buffers.
+  int                   _scratch_const_size;    // For temporary code buffers.
+  bool                  _in_scratch_emit_size;  // true when in scratch_emit_size.
 
  public:
   // Accessors
@@ -454,6 +539,12 @@ class Compile : public Phase {
                                                   _recent_alloc_obj = obj;
                                                 }
 
+  // Constant table
+  ConstantTable&   constant_table() { return _constant_table; }
+
+  MachConstantBaseNode*     mach_constant_base_node();
+  bool                  has_mach_constant_base_node() const { return _mach_constant_base_node != NULL; }
+
   // Handy undefined Node
   Node*             top() const                 { return _top; }
 
@@ -605,13 +696,16 @@ class Compile : public Phase {
   Dependencies*     dependencies()              { return env()->dependencies(); }
   static int        CompiledZap_count()         { return _CompiledZap_count; }
   BufferBlob*       scratch_buffer_blob()       { return _scratch_buffer_blob; }
-  void         init_scratch_buffer_blob();
+  void         init_scratch_buffer_blob(int const_size);
+  void        clear_scratch_buffer_blob();
   void          set_scratch_buffer_blob(BufferBlob* b) { _scratch_buffer_blob = b; }
   relocInfo*        scratch_locs_memory()       { return _scratch_locs_memory; }
   void          set_scratch_locs_memory(relocInfo* b)  { _scratch_locs_memory = b; }
 
   // emit to scratch blob, report resulting size
   uint              scratch_emit_size(const Node* n);
+  void       set_in_scratch_emit_size(bool x)   {        _in_scratch_emit_size = x; }
+  bool           in_scratch_emit_size() const   { return _in_scratch_emit_size;     }
 
   enum ScratchBufferBlob {
     MAX_inst_size       = 1024,
@@ -692,7 +786,7 @@ class Compile : public Phase {
   void Fill_buffer();
 
   // Determine which variable sized branches can be shortened
-  void Shorten_branches(Label *labels, int& code_size, int& reloc_size, int& stub_size, int& const_size);
+  void Shorten_branches(Label *labels, int& code_size, int& reloc_size, int& stub_size);
 
   // Compute the size of first NumberOfLoopInstrToAlign instructions
   // at the head of a loop.
diff --git a/src/share/vm/opto/gcm.cpp b/src/share/vm/opto/gcm.cpp
index 6a6a18e73..1a0116445 100644
--- a/src/share/vm/opto/gcm.cpp
+++ b/src/share/vm/opto/gcm.cpp
@@ -89,7 +89,7 @@ void PhaseCFG::replace_block_proj_ctrl( Node *n ) {
   assert(in0 != NULL, "Only control-dependent");
   const Node *p = in0->is_block_proj();
   if (p != NULL && p != n) {    // Control from a block projection?
-    assert(!n->pinned() || n->is_SafePointScalarObject(), "only SafePointScalarObject pinned node is expected here");
+    assert(!n->pinned() || n->is_MachConstantBase() || n->is_SafePointScalarObject(), "only pinned MachConstantBase or SafePointScalarObject node is expected here");
     // Find trailing Region
     Block *pb = _bbs[in0->_idx]; // Block-projection already has basic block
     uint j = 0;
diff --git a/src/share/vm/opto/machnode.cpp b/src/share/vm/opto/machnode.cpp
index 179116287..6242c66aa 100644
--- a/src/share/vm/opto/machnode.cpp
+++ b/src/share/vm/opto/machnode.cpp
@@ -489,6 +489,20 @@ void MachTypeNode::dump_spec(outputStream *st) const {
 }
 #endif
 
+
+//=============================================================================
+int MachConstantNode::constant_offset() {
+  int offset = _constant.offset();
+  // Bind the offset lazily.
+  if (offset == -1) {
+    Compile::ConstantTable& constant_table = Compile::current()->constant_table();
+    offset = constant_table.table_base_offset() + constant_table.find_offset(_constant);
+    _constant.set_offset(offset);
+  }
+  return offset;
+}
+
+
 //=============================================================================
 #ifndef PRODUCT
 void MachNullCheckNode::format( PhaseRegAlloc *ra_, outputStream *st ) const {
diff --git a/src/share/vm/opto/machnode.hpp b/src/share/vm/opto/machnode.hpp
index 633a9b143..d947bb871 100644
--- a/src/share/vm/opto/machnode.hpp
+++ b/src/share/vm/opto/machnode.hpp
@@ -231,9 +231,6 @@ public:
   // Return number of relocatable values contained in this instruction
   virtual int   reloc() const { return 0; }
 
-  // Return number of words used for double constants in this instruction
-  virtual int   const_size() const { return 0; }
-
   // Hash and compare over operands.  Used to do GVN on machine Nodes.
   virtual uint  hash() const;
   virtual uint  cmp( const Node &n ) const;
@@ -348,6 +345,65 @@ public:
 #endif
 };
 
+//------------------------------MachConstantBaseNode--------------------------
+// Machine node that represents the base address of the constant table.
+class MachConstantBaseNode : public MachIdealNode {
+public:
+  static const RegMask& _out_RegMask;  // We need the out_RegMask statically in MachConstantNode::in_RegMask().
+
+public:
+  MachConstantBaseNode() : MachIdealNode() {
+    init_class_id(Class_MachConstantBase);
+  }
+  virtual const class Type* bottom_type() const { return TypeRawPtr::NOTNULL; }
+  virtual uint ideal_reg() const { return Op_RegP; }
+  virtual uint oper_input_base() const { return 1; }
+
+  virtual void emit(CodeBuffer& cbuf, PhaseRegAlloc* ra_) const;
+  virtual uint size(PhaseRegAlloc* ra_) const;
+  virtual bool pinned() const { return UseRDPCForConstantTableBase; }
+
+  static const RegMask& static_out_RegMask() { return _out_RegMask; }
+  virtual const RegMask& out_RegMask() const { return static_out_RegMask(); }
+
+#ifndef PRODUCT
+  virtual const char* Name() const { return "MachConstantBaseNode"; }
+  virtual void format(PhaseRegAlloc*, outputStream* st) const;
+#endif
+};
+
+//------------------------------MachConstantNode-------------------------------
+// Machine node that holds a constant which is stored in the constant table.
+class MachConstantNode : public MachNode {
+protected:
+  Compile::Constant _constant;  // This node's constant.
+
+public:
+  MachConstantNode() : MachNode() {
+    init_class_id(Class_MachConstant);
+  }
+
+  virtual void eval_constant(Compile* C) {
+#ifdef ASSERT
+    tty->print("missing MachConstantNode eval_constant function: ");
+    dump();
+#endif
+    ShouldNotCallThis();
+  }
+
+  virtual const RegMask &in_RegMask(uint idx) const {
+    if (idx == mach_constant_base_node_input())
+      return MachConstantBaseNode::static_out_RegMask();
+    return MachNode::in_RegMask(idx);
+  }
+
+  // Input edge of MachConstantBaseNode.
+  uint mach_constant_base_node_input() const { return req() - 1; }
+
+  int  constant_offset();
+  int  constant_offset() const { return ((MachConstantNode*) this)->constant_offset(); }
+};
+
 //------------------------------MachUEPNode-----------------------------------
 // Machine Unvalidated Entry Point Node
 class MachUEPNode : public MachIdealNode {
diff --git a/src/share/vm/opto/matcher.hpp b/src/share/vm/opto/matcher.hpp
index 6e8f04c5c..c1627c60f 100644
--- a/src/share/vm/opto/matcher.hpp
+++ b/src/share/vm/opto/matcher.hpp
@@ -365,6 +365,10 @@ public:
   // registers?  True for Intel but false for most RISCs
   static const bool clone_shift_expressions;
 
+  // Should constant table entries be accessed with loads using
+  // absolute addressing?  True for x86 but false for most RISCs.
+  static const bool constant_table_absolute_addressing;
+
   static bool narrow_oop_use_complex_address();
 
   // Generate implicit null check for narrow oops if it can fold
diff --git a/src/share/vm/opto/node.hpp b/src/share/vm/opto/node.hpp
index c1aaf3600..67bcf90f2 100644
--- a/src/share/vm/opto/node.hpp
+++ b/src/share/vm/opto/node.hpp
@@ -81,6 +81,8 @@ class MachCallLeafNode;
 class MachCallNode;
 class MachCallRuntimeNode;
 class MachCallStaticJavaNode;
+class MachConstantBaseNode;
+class MachConstantNode;
 class MachIfNode;
 class MachNode;
 class MachNullCheckNode;
@@ -566,10 +568,12 @@ public:
               DEFINE_CLASS_ID(MachCallDynamicJava,  MachCallJava, 1)
             DEFINE_CLASS_ID(MachCallRuntime,      MachCall, 1)
               DEFINE_CLASS_ID(MachCallLeaf,         MachCallRuntime, 0)
-      DEFINE_CLASS_ID(MachSpillCopy, Mach, 1)
-      DEFINE_CLASS_ID(MachNullCheck, Mach, 2)
-      DEFINE_CLASS_ID(MachIf,        Mach, 3)
-      DEFINE_CLASS_ID(MachTemp,      Mach, 4)
+      DEFINE_CLASS_ID(MachSpillCopy,    Mach, 1)
+      DEFINE_CLASS_ID(MachNullCheck,    Mach, 2)
+      DEFINE_CLASS_ID(MachIf,           Mach, 3)
+      DEFINE_CLASS_ID(MachTemp,         Mach, 4)
+      DEFINE_CLASS_ID(MachConstantBase, Mach, 5)
+      DEFINE_CLASS_ID(MachConstant,     Mach, 6)
 
     DEFINE_CLASS_ID(Proj,  Node, 2)
       DEFINE_CLASS_ID(CatchProj, Proj, 0)
@@ -734,6 +738,8 @@ public:
   DEFINE_CLASS_QUERY(MachCallLeaf)
   DEFINE_CLASS_QUERY(MachCallRuntime)
   DEFINE_CLASS_QUERY(MachCallStaticJava)
+  DEFINE_CLASS_QUERY(MachConstantBase)
+  DEFINE_CLASS_QUERY(MachConstant)
   DEFINE_CLASS_QUERY(MachIf)
   DEFINE_CLASS_QUERY(MachNullCheck)
   DEFINE_CLASS_QUERY(MachReturn)
diff --git a/src/share/vm/opto/output.cpp b/src/share/vm/opto/output.cpp
index 015b8f69a..e2ddd974f 100644
--- a/src/share/vm/opto/output.cpp
+++ b/src/share/vm/opto/output.cpp
@@ -61,11 +61,6 @@ void Compile::Output() {
   // RootNode goes
   assert( _cfg->_broot->_nodes.size() == 0, "" );
 
-  // Initialize the space for the BufferBlob used to find and verify
-  // instruction size in MachNode::emit_size()
-  init_scratch_buffer_blob();
-  if (failing())  return; // Out of memory
-
   // The number of new nodes (mostly MachNop) is proportional to
   // the number of java calls and inner loops which are aligned.
   if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
@@ -333,7 +328,7 @@ void Compile::compute_loop_first_inst_sizes() {
 //----------------------Shorten_branches---------------------------------------
 // The architecture description provides short branch variants for some long
 // branch instructions. Replace eligible long branches with short branches.
-void Compile::Shorten_branches(Label *labels, int& code_size, int& reloc_size, int& stub_size, int& const_size) {
+void Compile::Shorten_branches(Label *labels, int& code_size, int& reloc_size, int& stub_size) {
 
   // fill in the nop array for bundling computations
   MachNode *_nop_list[Bundle::_nop_count];
@@ -353,12 +348,11 @@ void Compile::Shorten_branches(Label *labels, int& code_size, int& reloc_size, i
   // Size in bytes of all relocation entries, including those in local stubs.
   // Start with 2-bytes of reloc info for the unvalidated entry point
   reloc_size = 1;          // Number of relocation entries
-  const_size = 0;          // size of fp constants in words
 
   // Make three passes.  The first computes pessimistic blk_starts,
-  // relative jmp_end, reloc_size and const_size information.
-  // The second performs short branch substitution using the pessimistic
-  // sizing. The third inserts nops where needed.
+  // relative jmp_end and reloc_size information.  The second performs
+  // short branch substitution using the pessimistic sizing.  The
+  // third inserts nops where needed.
 
   Node *nj; // tmp
 
@@ -381,7 +375,6 @@ void Compile::Shorten_branches(Label *labels, int& code_size, int& reloc_size, i
         MachNode *mach = nj->as_Mach();
         blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
         reloc_size += mach->reloc();
-        const_size += mach->const_size();
         if( mach->is_MachCall() ) {
           MachCallNode *mcall = mach->as_MachCall();
           // This destination address is NOT PC-relative
@@ -398,10 +391,6 @@ void Compile::Shorten_branches(Label *labels, int& code_size, int& reloc_size, i
           if (min_offset_from_last_call == 0) {
             blk_size += nop_size;
           }
-        } else if (mach->ideal_Opcode() == Op_Jump) {
-          const_size += b->_num_succs; // Address table size
-          // The size is valid even for 64 bit since it is
-          // multiplied by 2*jintSize on this method exit.
         }
       }
       min_offset_from_last_call += inst_size;
@@ -562,10 +551,6 @@ void Compile::Shorten_branches(Label *labels, int& code_size, int& reloc_size, i
   // a relocation index.
   // The CodeBuffer will expand the locs array if this estimate is too low.
   reloc_size   *= 10 / sizeof(relocInfo);
-
-  // Adjust const_size to number of bytes
-  const_size   *= 2*jintSize; // both float and double take two words per entry
-
 }
 
 //------------------------------FillLocArray-----------------------------------
@@ -1102,10 +1087,39 @@ void Compile::Fill_buffer() {
     blk_labels[i].init();
   }
 
+  if (has_mach_constant_base_node()) {
+    // Fill the constant table.
+    // Note:  This must happen before Shorten_branches.
+    for (i = 0; i < _cfg->_num_blocks; i++) {
+      Block* b = _cfg->_blocks[i];
+
+      for (uint j = 0; j < b->_nodes.size(); j++) {
+        Node* n = b->_nodes[j];
+
+        // If the node is a MachConstantNode evaluate the constant
+        // value section.
+        if (n->is_MachConstant()) {
+          MachConstantNode* machcon = n->as_MachConstant();
+          machcon->eval_constant(C);
+        }
+      }
+    }
+
+    // Calculate the offsets of the constants and the size of the
+    // constant table (including the padding to the next section).
+    constant_table().calculate_offsets_and_size();
+    const_req = constant_table().size();
+  }
+
+  // Initialize the space for the BufferBlob used to find and verify
+  // instruction size in MachNode::emit_size()
+  init_scratch_buffer_blob(const_req);
+  if (failing())  return; // Out of memory
+
   // If this machine supports different size branch offsets, then pre-compute
   // the length of the blocks
   if( _matcher->is_short_branch_offset(-1, 0) ) {
-    Shorten_branches(blk_labels, code_req, locs_req, stub_req, const_req);
+    Shorten_branches(blk_labels, code_req, locs_req, stub_req);
     labels_not_set = false;
   }
 
@@ -1121,12 +1135,12 @@ void Compile::Fill_buffer() {
     code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10;  // force expansion
 
   int total_req =
+    const_req +
     code_req +
     pad_req +
     stub_req +
     exception_handler_req +
-    deopt_handler_req +              // deopt handler
-    const_req;
+    deopt_handler_req;               // deopt handler
 
   if (has_method_handle_invokes())
     total_req += deopt_handler_req;  // deopt MH handler
@@ -1180,6 +1194,11 @@ void Compile::Fill_buffer() {
 
   NonSafepointEmitter non_safepoints(this);  // emit non-safepoints lazily
 
+  // Emit the constant table.
+  if (has_mach_constant_base_node()) {
+    constant_table().emit(*cb);
+  }
+
   // ------------------
   // Now fill in the code buffer
   Node *delay_slot = NULL;
@@ -1196,12 +1215,13 @@ void Compile::Fill_buffer() {
       cb->flush_bundle(true);
 
     // Define the label at the beginning of the basic block
-    if( labels_not_set )
-      MacroAssembler(cb).bind( blk_labels[b->_pre_order] );
-
-    else
-      assert( blk_labels[b->_pre_order].loc_pos() == cb->insts_size(),
-              "label position does not match code offset" );
+    if (labels_not_set) {
+      MacroAssembler(cb).bind(blk_labels[b->_pre_order]);
+    } else {
+      assert(blk_labels[b->_pre_order].loc_pos() == cb->insts_size(),
+             err_msg("label position does not match code offset: %d != %d",
+                     blk_labels[b->_pre_order].loc_pos(), cb->insts_size()));
+    }
 
     uint last_inst = b->_nodes.size();
 
@@ -1718,9 +1738,17 @@ void Compile::ScheduleAndBundle() {
   // Create a data structure for all the scheduling information
   Scheduling scheduling(Thread::current()->resource_area(), *this);
 
+  // Initialize the space for the BufferBlob used to find and verify
+  // instruction size in MachNode::emit_size()
+  init_scratch_buffer_blob(MAX_const_size);
+  if (failing())  return;  // Out of memory
+
   // Walk backwards over each basic block, computing the needed alignment
   // Walk over all the basic blocks
   scheduling.DoScheduling();
+
+  // Clear the BufferBlob used for scheduling.
+  clear_scratch_buffer_blob();
 }
 
 //------------------------------ComputeLocalLatenciesForward-------------------
diff --git a/src/share/vm/opto/postaloc.cpp b/src/share/vm/opto/postaloc.cpp
index c901f611f..897d5102c 100644
--- a/src/share/vm/opto/postaloc.cpp
+++ b/src/share/vm/opto/postaloc.cpp
@@ -200,6 +200,19 @@ int PhaseChaitin::elide_copy( Node *n, int k, Block *current_block, Node_List &v
   // then reloaded BUT survives in a register the whole way.
   Node *val = skip_copies(n->in(k));
 
+  if (val == x && nk_idx != 0 &&
+      regnd[nk_reg] != NULL && regnd[nk_reg] != x &&
+      n2lidx(x) == n2lidx(regnd[nk_reg])) {
+    // When rematerialzing nodes and stretching lifetimes, the
+    // allocator will reuse the original def for multidef LRG instead
+    // of the current reaching def because it can't know it's safe to
+    // do so.  After allocation completes if they are in the same LRG
+    // then it should use the current reaching def instead.
+    n->set_req(k, regnd[nk_reg]);
+    blk_adjust += yank_if_dead(val, current_block, &value, &regnd);
+    val = skip_copies(n->in(k));
+  }
+
   if( val == x ) return blk_adjust; // No progress?
 
   bool single = is_single_register(val->ideal_reg());